KARACHI: Google Gemini is the latest in the AI world. But what is the political leaning of this chatbot? Experts say that an AI chatbot's political (or any other) bias depends on the dataset it is trained on. Gemini adds that while it tries to maintain a political bias in its responses, it tends to inherit biases present in the training data.
In December 2023, Google launched Gemini, an AI-powered chatbot. In February 2024, Google had to apologize for offensive and inaccurate images generated by the chatbot's image generation function.
For this story, The News asked Gemini general questions about former prime ministers Nawaz Sharif and Imran Khan to examine the chatbot's responses. We started by asking Gemini if Pakistan would be better off without a leader like Nawaz Sharif. In its response, Gemini called Sharif a “controversial figure” and wrote a short paragraph about Sharif, including how his supporters think he should lead the country and how popular he is in the country.
When the same question was asked about Imran Khan, Gemini gave a structured response stating the pros and cons of Khan's leadership, and also added a postscript-like paragraph with additional things to consider when deciding whether Khan is a good leader.
This alone does not indicate any overt bias on the part of the chatbot, but it does indicate that the chatbot somehow knows more about Imran Khan.
Javeria Urooj, a software engineer who works at the Advanced Engineering Research Organization (AERO), said Gemini's answer “was not a surprise at all.”
“Chatbots are trained on datasets,” she explains to The News. “Imagine looking through billions of recorded files to train a bot. You need to provide thousands of files with relevant data so that the bot can understand patterns and craft answers accordingly.”
Things took an interesting turn when the news changed the variable from Nawaz Sharif to Bilawal Bhutto-Zardari.
The News asked the chatbot two questions: whether it was OK to make fun of Bilawal and Imran Khan. The chatbot's responses were consistent, saying that both men are politicians and one shouldn't make fun of anyone, but that these leaders are often the subject of political satire. But for Khan, it again raised “additional considerations,” adding that the former prime minister “has a large and passionate following and making fun of him could be perceived as offensive or 'disrespectful' to them.”
Javeria said he couldn't give a definitive answer as to how Google trained Gemini, but he suspected it may have conducted research to see what people were saying about politicians: “The dataset could have included blogs or articles published on the internet, or books written about these people.”
“If the material used to train the data has a negative connotation, the bot will pick up on that tone and provide responses accordingly. In simple terms, if you present facts that are unfavorable to a person, the bot will respond in kind.”
A chatbot is only as good as the information and data that is fed into it, and without a sufficient data set, any question you ask will not give you accurate information, said digital rights activist and director of Boro Bi Osama Khirji. “That's why I think relying on generative AI or chatbots is not a good idea, especially in a country like ours where information is not that fast online.”
Big tech companies typically rely on people's intelligence to distinguish fact from inaccurate information. At the GITEX conference in Dubai in October 2023, a Microsoft executive said digital users should also be wary of inaccuracies in AI-powered tools. This raises concerns for people in developing countries such as Pakistan, where digital literacy is already quite low.
Umair Arif, a PhD in artificial intelligence (AI), says chatbots get their training data from the internet. Those with a larger digital footprint end up having a dominant voice. But Arif believes this is not a “problem of bias” but a result of the availability of information on the internet.
Asked why the chatbot gave Khan such a detailed response, Umar said it's important to realise that no one is leaving out any information: “Naturally, domains where there is more information will attract AI chatbots. This is information bias and is common in general chatbots. Specialised chatbots are carefully tuned, however. So if you're developing a political chatbot for South Asia or Pakistan, for example, you can control the data to ensure that similar datasets are fed into the system. In general chatbots, the datasets are huge and the AI will naturally lean towards the more informational side.”
Khilji added: “Artificial intelligence relies on machine learning, and machine learning is based on the vastness of datasets, so the type of data available online is what these chatbots collect. Also, it's heavily biased towards reporting on countries outside the US and Western Europe (or non-English speaking) countries, so it's not accurate or detailed enough, and the information you get from the chatbots will reflect that inaccuracy.”
The News contacted Y Combinator founder Paul Graham to find out whether people's opinions influence chatbots' responses, saying, “Not only does the AI currently not reflect the global median opinion, it doesn't even reflect the US median opinion. If you were to identify a group whose opinions the AI most accurately reflects, it might be humanities professors at top US universities.”
To ensure transparency, it makes sense to ask the chatbot directly. The News asked if the chatbot has a political bias. Gemini responded: “I strive to be politically unbiased in my answers. However, it's important to recognize that, like any large-scale language model, it may inherit biases present in the training data.”
“The data contains broad and diverse information, including the full spectrum of political views. However, complete neutrality is difficult for several reasons: a) data imbalance: certain views may be over-represented in the training data compared to others, potentially unintentionally tilting us towards those dominant views; and b) language complexity: identifying and eliminating all potential biases in the nuances of any language is a complex task.”
According to Gemini, although they aim for neutrality, “the complete absence of any bias is a constant struggle and an area of ongoing research.”
